Sign in Start free
CONTENT WRITING

Fact-check pass on a draft

Use before publishing anything that contains numbers, dates, claims or attributions you did not personally verify.

fact-check-pass.md
Download .md 
You are a fact-checker reviewing a draft before publication. You are not editing style.

Draft:
{{DRAFT}}
Sources I have already supplied to the writer:
{{SUPPLIED_SOURCES}}
Publication date context: {{PUBLICATION_DATE}}

Produce one table. Columns: Claim (quoted) | Type (statistic, date, attribution, causal claim, superlative, product capability) | Supported by supplied sources? (yes/partial/no) | Risk if wrong (high/medium/low) | What to do.

Then two short lists:
A. Claims you cannot verify from what I gave you and that need a primary source. Name the kind of source that would settle it.
B. Claims that are probably true but are stated too strongly, with a suggested weaker wording.

Rules:
- Do not assert a fact is correct from your own memory. Your job is to sort claims by verifiability, not to supply the verification.
- Flag any number with no date attached as a risk, since figures decay.
- Flag superlatives such as leading, fastest or most popular as requiring evidence.
- Quote claims exactly. Do not paraphrase them into the table.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

  • {{DRAFT}}
  • {{SUPPLIED_SOURCES}}
  • {{PUBLICATION_DATE}}

Getting a better result

  1. Paste the source documents in full where you can, so the supported column means something.
  2. Sort by the risk column and fix the high rows first if time is short.
  3. Anything in list A that you cannot source should be cut, not softened.

Questions about this prompt

When is this worth running rather than checking the claims myself?

Before publishing anything with numbers, dates, attributions or capability claims you did not verify. It does not replace checking. The prompt is told never to assert a fact from its own memory, so it sorts claims by verifiability and hands you a queue, which is the honest division of labour with a model.

What do I have to supply for the table to mean anything?

The draft, the sources you gave the writer pasted in full wherever you can, and the publication date context. Links instead of source text make the supported column meaningless, since the model cannot open them. The date decides which figures count as stale rather than merely undated.

What does the check produce?

One table classifying every quoted claim by type, support, risk and action, then list A of claims needing a primary source and list B of claims that are probably true but stated too strongly. Sort by risk and clear the high rows first. List B is usually minutes of work for real exposure.

What is the tempting shortcut to avoid?

Softening a list A claim instead of cutting it. Rewording an unsourceable statistic leaves the same problem in the piece with less to check it against. Expect superlatives such as leading or fastest to be flagged as well; those need evidence rather than a hedge in front of them.